Liquid AI reveals 8B-A1B MoE trained on 38T
Recorded: May 29, 2026, 7:03 p.m.
| Original | Summarized |
LFM2.5-8B-A1B: an Even Better on-Device Mixture-of-Experts | Liquid AI Products Liquid Foundation ModelsThe world’s most efficient models LEAPLFM customization and deployment platform Liquid ApolloPrivate on-device AI for everyone Pricing & LicensingSolutions Enterprise SolutionsUnlock your business with the most efficient AI on the market.Startup SolutionsExpert support in creating the LFM that’s right for your startup.AutomotiveConsumer ElectronicsEcommerceFinanceHealthcare and Life SciencesIndustrial & RoboticsResources BlogCase StudiesDemosDocumentationDocs, use cases, and cookbooksDeveloper CommunityHackathonsCompany AboutOur PartnersNewsroomCareers Get Liquid Download ModelsLEAPHuggingFaceAmazon BedrockCustomize & Deploy ModelsLEAPPlay with ModelsLiquid PlaygroundLiquid Apollo on App StoreLiquid Apollo on Google PlayRequest a DemoTry Liquid Products OUR PRODUCTSLiquid Foundation ModelsThe world’s most efficient modelsLEAPLFM customization and deployment platform Liquid ApolloPrivate on-device AI for everyone PRICINGPricing & LicensingSolutions INDUSTRIESAutomotiveConsumer ElectronicsEcommerceFinanceHealthcare & Life SciencesIndustrial & RoboticsEFFICIENCY MEETS CAPABILITYEnterprise SolutionsUnlock your business with the most efficient AI on the market.Startup SolutionsExpert support in creating the LFM that’s right for your startup.ResearchResources BlogCase StudiesDemosDocumentationDocs, use cases, and cookbooksDeveloper CommunityHackathonscompany AboutOur PartnersCareers Newsroom EnglishJapaneseModelsLFM2.5-8B-A1B: an Even Better on-Device Mixture-of-ExpertsAuthorsLiquid AIPublishedMay 28, 2026OverviewShare: Today, we're releasing LFM2.5-8B-A1B, an edge model built for fast, reliable tool calling on consumer hardware.It builds on our LFM2-8B-A1B release from October 2025, with an expanded 128K context window, scaled-up pretraining (from 12T to 38T tokens), and large-scale reinforcement learning. We also doubled its vocabulary to improve tokenization efficiency for non-Latin languages. The result is a model that chains tool calls, achieves tasks, and fits comfortably even on an entry-level laptop.The base (LFM2.5-8B-A1B-Base) and post-trained (LFM2.5-8B-A1B) models are available today on Hugging Face and our Playground. Check out our docs on how to run and fine-tune them locally.*AA-Omniscience Index (higher is better) rewards correct answers and penalizes hallucinations. Scores range from -100 to 100. See more results on Artificial Analysis.HighlightsOn-device personal assistant. Designed to power real-life applications, chaining tool calls, and following complex instructions on all devices.Compressed performance. Competitive with much larger dense and MoE models on instruction following and agentic tasks.Unmatched throughput. Fastest in its size class on both CPU and GPU inference, with day-one support for llama.cpp, MLX, vLLM, and SGLang.What changed since LFM2-8B-A1BCompared to LFM2-8B-A1B, this new version expands the context window from 32,768 to 128,000 tokens. This allows the model to process longer documents and reason for longer. Its vocabulary size was also scaled up from 65,536 to 128,000 to tokenize non-Latin scripts more efficiently. We see particularly strong compression gains in Hindi, Thai, Vietnamese, Indonesian, and Arabic. The rest of the architecture follows the same combination of MoE, GQA, and gated short convolution blocks as LFM2-8B-A1B, as shown in the following figure.Unlike its predecessor, LFM2.5-8B-A1B is a reasoning-only model, producing an explicit chain of thought before its final answer. We adopted this strategy because MoE models generally run in compute-bound settings, where a smaller number of active parameters makes each reasoning token cheap. This provides a significant quality boost without compromising speed.Thanks to reasoning and scaled-up training, this new version performs significantly better: Benchmark LFM2-8B-A1B LFM2.5-8B-A1B Δ AA-Omniscience Index -78.42 -24.70 +53.62 AA-Omniscience Accuracy 7.33 8.67 +1.34 AA-Omniscience Non-Hallucination Rate 7.46 63.47 +56.01 IFEval 79.44 91.84 +12.40 IFBench 26.00 56.47 +30.47 Multi-IF 58.54 79.93 +21.39 MATH500 74.80 88.76 +13.96 AIME25 20.00 42.53 +22.53 BFCLv3 45.07 64.36 +19.29 BFCLv4 25.52 48.50 +22.98 Tau² Telecom 13.60 88.07 +74.47 Tau² Retail 7.02 39.82 +32.80 Training highlightsTokenizer expansion. LFM2-8B-A1B was originally trained with a 65K BPE tokenizer optimized for our initial language coverage. To better support non-Latin scripts in LFM2.5, we doubled the vocabulary to 128K by extending the existing tokenizer in place rather than retraining the model from scratch.. We continued BPE merge training from the original merges on a multilingual corpus, which keeps most existing token IDs as identity mappings and makes every new token decompose deterministically into a sequence of original sub-tokens. We initialize the new embedding rows as the mean of their sub-token decompositions and copy the shared rows unchanged. We then recover quality through a brief two-stage adaptation: embedding-only training, followed by full-model continued pretraining.The table below reports chars/token, roughly how much text each token carries: higher is better, and the new tokenizer is more efficient in all 16 languages Tokenizer Arabic (ar) Old tokenizer 2.239 New tokenizer 3.107 Improvement +38.8% Context extension. We first extended the context window to 32K through a 2T token midtraining phase focused on reasoning, math, tool-use, and longer documents. We then extended the context to 128K by increasing the RoPE base θ and running an additional 400B token midtraining stage focused on long-document and long-trajectory data.Doom loops. We added a targeted preference optimization stage to reduce doom loops in long reasoning traces. This stage identifies tokens that tend to trigger looping behavior in specific contexts, then redistributes probability mass toward plausible alternatives, while leaving the rest of the next-token distribution largely intact. During RL, we also added a lightweight shaping reward that discourages excessive use of common loop-inducing restart words like “Wait…”. We'll share more details on the full pipeline, objective, and empirical results in a dedicated blog post.Hallucinations. Because of their small number of parameters, edge models have a limited knowledge capacity, which leads to more hallucinations. To mitigate hallucinations, we added a targeted RL stage that uses an avg@k-based reward over a diverse knowledge dataset. The goal is to reinforce abstention on queries beyond reliable knowledge while preserving existing knowledge. This produces a sharper knowledge boundary and clearer expression of uncertainty.BenchmarksWe evaluated LFM2.5-8B-A1B across benchmarks covering knowledge, instruction following, math, and agentic workflows. The model is competitive with both dense alternatives with a similar total number of parameters and much larger MoEs. Model Parameters AA-Omniscience Index Accuracy Non-Hallucination IFEval IFBench Multi-IF LFM2.5-8B-A1B 8B/A1B Granite-4.0-H-Tiny 7B/A1B Qwen3.5-4B 4B Qwen3-30B-A3B-Thinking-2507 30.5B/3.3B Gemma-4-E2B-IT 5.1B Gemma-4-E4B-IT 8B Gemma-4-26B-A4B-IT 26B/4B gpt-oss-20b 21B/3.6B The avg@k-based reward enables LFM2.5-8B-A1B to achieve a significantly lower hallucination rate while maintaining reasonable accuracy. It also leads on instruction following benchmarks, matching bigger MoEs like Gemma 4-26B at a fraction of the active parameter count.Math and agentic workflows Model Parameters MATH500 AIME25 AIME26 BFCLv3 BFCLv4 Tau² Telecom Tau² Retail LFM2.5-8B-A1B 8B/A1B Granite-4.0-H-Tiny 7B/A1B Qwen3.5-4B 4B Qwen3-30B-A3B-Thinking-2507 30.5B/3.3B Gemma-4-E2B-IT 5.1B Gemma-4-E4B-IT 8B Gemma-4-26B-A4B-IT 26B/4B gpt-oss-20b 21B/3.6B On agentic benchmarks, LFM2.5-8B-A1B is competitive with bigger models and particularly strong on Tau2-Telecom. As agentic harnesses are becoming the main way to consume models, LFM2.5-8B-A1B is a first step towards powering on-device, fully private agents.Sparse Inference, EverywhereLFM2.5-8B-A1B ships with day-one support across the inference ecosystem:LEAP — Liquid's Edge AI Platform for iOS and Android deploymentllama.cpp — GGUF checkpoints for efficient edge inferenceMLX — Optimized inference for Apple SiliconvLLM — GPU-accelerated serving for production throughputSGLang — GPU-accelerated serving for production throughputONNX — Cross-platform inference across diverse acceleratorsCPU inference. LFM2.5-8B-A1B ships with day-one llama.cpp support and runs on everyday consumer hardware.On both laptop-class chips, it is the fastest model we tested at reading in prompts and generating answers, decoding 253 tokens/s on an M5 Max and 146 on a Ryzen AI Max+ 395 while staying under 6 GB. It even holds ~30 tokens/s on a phone, so a capable assistant runs instantly and privately on your own device.GPU inference. We support inference via vLLM and SGLang via active contributions to these codebases. We measure output throughput (total output tokens divided by wall time) on a single NVIDIA H100 SXM5 GPU using a sustained-load setting: at each concurrency level, we continuously maintain the target number of in-flight requests, replacing each completed request immediately.We benchmark each model with SGLang 0.5.12, 1,024 input tokens, up to 256 output tokens, in BF16, averaging 3 runs per concurrency level. LFM2.5-8B-A1B is the fastest model in its size class, reaching 18.5K output tokens per second at high concurrency, over 1.6B tokens per day on a single H100.Local Cowork: see it runOur open-source desktop agent demo, LocalCowork, now runs on LFM2.5-8B-A1B. The setup is the same one we used for LFM2-24B-A2B demo in March: a single laptop, 67 tools across 13 MCP servers, no cloud, no API keys, no data leaving the machine. Tool selection is faster and noticeably more reliable across the same tool menu.The point of the demo is not the individual tools. It is that the tool-dispatch loop feels interactive on consumer hardware: ask, propose, confirm, run, repeat, all in well under a second per dispatch, with full audit trails and your data never leaving the device.Get StartedWith LFM2.5, we're delivering on our vision of AI that runs anywhere. These models are:Open-weight — Download, fine-tune, and deploy without restrictionsFast from day one — Native support for llama.cpp, MLX, vLLM, SGLang across Apple, AMD, Intel, Qualcomm, and Nvidia hardwareA complete family — From base models for customization to specialized audio and vision variants, one architecture covers diverse use casesThe on-device agentic future starts here. We can't wait to see what you build. Try on Playground Download on Hugging Face Read the Docs Citation Liquid AI, “LFM2.5-8B-A1B: Personal Assistant On Your Laptop,” Ready to experience AI?Power your business, workflows, and engineers with Liquid AI.Try Liquid Talk to our teamStay up to date with LiquidThank you! Oops! Something went wrong while submitting the form. Get Liquid Apollo for freeVibe check LFMs directly on your phone. SolutionsEnterprise SolutionsStart Up SolutionsAutomotiveConsumer ElectronicsEcommerceFinancial ServicesHealthcare & Life SciencesIndustrial & RoboticsCase StudiesTech & ProductsLiquid Foundation ModelsLEAPLiquid ApolloResearchPricing & LicensingDeveloper ResourcesDemosDocumentationDeveloper CommunityHackathonsCompanyAboutBlogPressCareers Contact UsShoutoutsLegalLFM LicensePrivacy PolicyPrivacy ChoicesTerms & ConditionsGet Liquid Apollo for freeVibe check LFMs directly on your phone. © {year}, Liquid AI, Inc. All rights reserved. 314 Main St, Cambridge, MA 02142We use cookies to enhance your browsing experience and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. To control your cookie settings, choose "Customize".Reject allAccept all Manage your preferencesWe use cookies to enhance your browsing experience and analyze our traffic. By clicking "Accept All", you consent to our use of cookies.Learn more about our Privacy PolicyEssential cookies requiredMarketing cookiesAnalytics cookiesReject allAccept all |
Liquid AI introduced LFM2.5-8B-A1B, an edge model specifically engineered for fast and reliable tool calling on consumer hardware. This advancement builds upon the LFM2-8B-A1B, incorporating significant architectural and training enhancements to create a model capable of operating effectively on local devices. Key features of the new model include an expanded 128,000 token context window, scaled pretraining from 12 trillion to 38 trillion tokens, and large-scale reinforcement learning integration. The model is designed to function as an on-device personal assistant, adept at chaining tool calls and following intricate instructions across various devices. The development focused on optimizing performance while maintaining high capability. LFM2.5-8B-A1B is a reasoning-only model, which facilitates efficiency by ensuring that fewer active parameters are computationally cheap for each reasoning token, leading to a notable quality improvement. To enhance robustness and accuracy, the model incorporates specific training stages aimed at mitigating hallucinations, utilizing an average at k based reward over diverse knowledge datasets to reinforce the model's tendency to abstain when knowledge is uncertain, thereby establishing a sharper knowledge boundary. A critical aspect of this release was the expansion of the tokenizer to a vocabulary size of 128,000, which was achieved by extending the existing tokenizer and continuing BPE merge training from the original multilingual corpus. This expansion significantly improves tokenization efficiency, particularly for non-Latin scripts, leading to substantial compression gains across numerous languages such as Hindi, Thai, Vietnamese, Indonesian, and Arabic. This efficiency gain is reflected in the token-to-character ratios across different languages, demonstrating superior token utilization. The context extension was achieved through a multi-stage pretraining process. The model first expanded the context window to 32,000 tokens during a midtraining phase focused on reasoning, mathematics, tool use, and long documents. This was subsequently extended to 128,000 tokens through an additional 400 billion token midtraining stage focused on long-document and long-trajectory data. Furthermore, the training included targeted steps to reduce undesirable behaviors, such as doom loops in long reasoning traces, by implementing a preference optimization stage that redirects probability mass toward plausible alternatives. The model’s deployment capability is supported by sparse inference strategies and broad hardware compatibility. LFM2.5-8B-A1B offers native support for various efficient inference frameworks, including llama.cpp for GGUF checkpoints, MLX for Apple Silicon optimization, vLLM for GPU-accelerated serving, SGLang for GPU server execution, and ONNX for cross-platform inference. This allows the model to run efficiently on everyday consumer hardware, including laptops, with speeds benchmarked against other models, showcasing exceptional throughput on both CPU and GPU. The model’s agentic potential is demonstrated through demonstrations like LocalCowork, an open-source desktop agent running entirely locally on a single laptop without reliance on the cloud or external API keys. This showcases the model's ability to facilitate interactive tool-dispatch loops with high reliability and speed on consumer hardware, positioning LFM2.5-8B-A1B as a foundation for on-device, fully private agents. The model is released as open-weight, allowing users to download, fine-tune, and deploy it without restrictions, supported by a complete family of models catering to diverse use cases. |